Federated Learning with Heterogeneous Labels and Models for Mobile Activity Monitoring

HARE: Unifying the Human Activity Recognition Engineering Workflow

“…Privacy is a major concern in the field of HAR, where the collection and processing of personal data are an integral part of the process [35]. Several existing approaches, such as federated learning, differential privacy, and homomorphic encryption, have been proposed to address this issue and enable the training of HAR models on personal devices without compromising privacy [36,37]. However, challenges remain in ensuring the fidelity and security of the collected data.…”

Robust Convergence in Federated Learning through Label-wise Clustering

“…Xie et al [11] presented a new approach called multi-center FL clustering, by constructing multiple global models and assigning them to the nearest local models. Many studies [17], [54], [55] have implemented auxiliary public dataset to minimize the biased distribution of class labels of existing clients, which accommodates the improvement of training performance in FL. Li et al [16] curated an additional public dataset that could assist the training among the local models and adopted transfer learning with knowledge distillation in the FL network.…”

“…These shortcomings imply that distance-based local weight clustering is an approach that is intractable to explain, which gives us an inspiration to design an algorithm that clusters through the given local labels. Although past several studies has emphasized that biased labels are major defect that perturbs the functionality under FL settings, they suggested alternatives [17], [54][55][56], [60], [61] such as implementing public dataset. To the best of our knowledge, there were no published articles that selects local models through evaluating their cor-responding labels with statistical metrics without using auxiliary dataset on FL framework.…”

“…However, acquiring IID (Independently and Identically Distributed) data in a designated local machine is not guaranteed in a realworld distributed environment, as most of the existing systems tend to be biased and skewed due to intrinsic propensity and external factors that influence the circumstances. Thus, measures in order to upgrade the FL performance were introduced to successfully mitigate the unstable impact of aggregating non-IID data through approaches such as utilizing learning paradigms [4][5][6][7][8][9], clustering [10][11][12][13][14], knowledge distillation [15][16][17][18], stochastic approach [11], [19][20], Bayesian theorem [21][22][23][24], etc. Initially, FL architecture was specifically designed in order to accommodate the effective training among the distributed conditions without sending the local data itself for security purposes [1], [2], [25], thus many researches were conducted on FL to achieve higher privacy such as detecting adversaries [12], [26].…”

“…Moreover, FL has been applied in the network domain as well, in order to construct efficient communication system architecture among heterogeneous local models to reduce the overhead [27], [28]. Likewise, FL was suggested as a viable methodology that effectively encompasses the geographically dispersed local models that were trained independently, which lead the diverse domains to implement the FL architecture such as in healthcare [17], [29][30][31][32][33], IoT [34][35][36][37] or Security [38][39][40][41][42][43][44].…”